Joint modeling of bivariate longitudinal and survival data in spouse pairs

Joint modeling of longitudinal and survival data has become increasingly useful for analyzing clinical trials data. Recent multivariate joint models relate one or more longitudinal outcomes to one or more failure times (e.g., competing risks) in the same subject. We consider a case where longitudinal and survival outcomes are measured in subject pairs (e.g., married couples). In this dissertation, we propose a joint model incorporating within-pair correlations, both in the longitudinal and survival processes. We use a bivariate linear mixed-effects model for the longitudinal process, where the random effects are used to model the temporal correlation among longitudinal outcomes and the correlation between different outcomes. For the survival process, we incorporate a gamma frailty into a Weibull proportional hazards model to account for the correlation between survival times within pairs. The two sub-models are then linked through the shared random effects, where the longitudinal and survival processes are conditionally independent given the random effects. Parameter estimates are obtained by maximizing the joint likelihood for the bivariate longitudinal and bivariate survival data using the EM algorithm.\ud The proposed methodology is applied to the spouse data from the Cardiovascular Health Study (CHS) to investigate the association of both longitudinal depression scores and survival times between husbands and wives, and to quantify the association of mortality and longitudinal depression with other covariates in husbands and wives after accounting for the within-spouse correlation. Public Heath Significance: Spouse studies seek to reveal the importance of both environmental and genetic influences on individuals. The analysis of such information is useful in assessing long term health effects in spouse pairs and/or individuals living together. The methodology we propose provides a valid statistical inference on the association of longitudinal measurements and the time-to-events among paired subjects. This methodology will contribute to the analysis of public health studies by ensuring that proper prediction and inference are made when pairs of individuals are measured longitudinally